<div class="gmail_quote">On Tue, Aug 31, 2010 at 7:20 PM, Ian Romanick <span dir="ltr"><<a href="mailto:idr@freedesktop.org">idr@freedesktop.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin: 0pt 0pt 0pt 0.8ex; border-left: 1px solid rgb(204, 204, 204); padding-left: 1ex;"> <div class="im">-----BEGIN PGP SIGNED MESSAGE-----<br> Hash: SHA1<br> <br> </div><div class="im">Marek Ol?ák wrote:<br> > On Tue, Aug 31, 2010 at 2:14 AM, Ian Romanick <<a href="mailto:idr@freedesktop.org">idr@freedesktop.org</a><br> </div><div class="im">> <mailto:<a href="mailto:idr@freedesktop.org">idr@freedesktop.org</a>>> wrote:<br> ><br> > While I was trying to get one of the Humus demos working today, it<br> > occurred to me that we can possibly do better than<br> > ir_vec_index_to_cond_assign to lower variable indexing of vectors. In<br> > addition to using conditional assignment, we can also use a dot-product<br> > to pick a single element out of a vector. The variable index operation<br> > becomes:<br> ><br> > const vec4 gl_vec_selector[4] =<br> > vec4[4](vec4(1.0, 0.0, 0.0, 0.0),<br> > vec4(0.0, 1.0, 0.0, 0.0),<br> > vec4(0.0, 0.0, 1.0, 0.0),<br> > vec4(0.0, 0.0, 0.0, 1.0));<br> ><br> > ...<br> ><br> > float f = dot(v, gl_vec_selector[i]);<br> <br> </div>[snip]<br> <div class="im"><br> > Neither r300 nor r500 supports the ARL opcode in fragment shaders (it's<br> <br> </div>Meh. I always forget about that asymmetry.<br> <div class="im"><br> > a D3D10 feature), which kind of makes this optimization a no-go. I<br> > suggest using SEQ instead:<br> ><br> > bvec4 selector = equal(vec4(i), vec4(0,1,2,3));<br> > float f = dot(v, vec4(selector));<br> ><br> > which should end up being just SEQ followed by DP4.<br> <br> </div>SEQ isn't part of the ARB_fragment_program / ARB_vertex_program<br> instruction set either. Does R300 support that? I won't be surprised<br> if i915 doesn't. Of course, it doesn't support the ARL-based<br> optimization either.<br></blockquote><div><br>R500 vertex shaders support SEQ natively.<br><br>For R300 vertex shaders, SEQ is lowered to a sequence of opcodes SGE, SGE, MUL. (because CMP is unsupported in hw)<br><br>For R300-R500 fragment shaders, SEQ is lowered to a sequence of opcodes ADD, CMP. (because SGE is unsupported in hw)<br> <br>SEQ is probably the best "high-level" instruction here, but I am ok with anything other than ARL.<br><br>Marek<br><br></div><blockquote class="gmail_quote" style="margin: 0pt 0pt 0pt 0.8ex; border-left: 1px solid rgb(204, 204, 204); padding-left: 1ex;"> <br> SEQ is itself lowered to a sequence of instructions:<br> <br> SGE t0, i.xxxx, { 0, 1, 2, 3};<br> SGE t1, i.xxxx, {-0, -1, -2, -3};<br> MUL selector, t0, t1;<br> DP4 result, v, selector;<br> <br> So, that probably still would be better than the mess that gets<br> generated today.<br> <div class="im">-----BEGIN PGP SIGNATURE-----<br> Version: GnuPG v1.4.10 (GNU/Linux)<br> Comment: Using GnuPG with Mozilla - <a href="http://enigmail.mozdev.org/" target="_blank">http://enigmail.mozdev.org/</a><br> <br> </div>iEYEARECAAYFAkx9OfMACgkQX1gOwKyEAw+HlQCfZCqRdGwn3x9k06/qReehqMSf<br> nCoAnjKMPoUlb/4QXeHW0EYOozlLpQn2<br> =X6R+<br> -----END PGP SIGNATURE-----<br> </blockquote></div><br>